Mathematical modelling offers a cost-effective alternative to designing Li- ion batteries (LiBs) compared to physical prototyping, which is both expensive and time-consuming. However, the efficacy of mathematical modelling of LiBs is predicated on accurate battery parameters. Parameterisation for LiB models can be challenging, often requiring precise measurements that are difficult or costly to obtain. In this thesis, we develop a fast and efficient approach for inferring the parameter values of LiB models using neural network surrogate (NNS) models. NNS models offer advantages in terms of speed, analytical functionality, and differentiability, making them well-suited for parallel processing. Developing NNS counterparts to LiB models provides an effective alternative for approximating these models and inferring their parameters efficiently. We develop different NNS models designed to predict battery performance. First, we present a NNS that is trained by data generated by solving the Doyle-Fuller-Newman (DFN) model using a classical solver and pre-specified LiB properties. Second, we introduce a NNS that is trained on an arbitrary range of parameters for the Th´evenin model. Third, we present a NNS that is trained by the data generated by the Th´evenin model, subject to a time-varying current. Then, we propose a Bayesian inference approach that utilises the trained NNS model of the Th´evenin model, subject to a time-varying current, to parameterise the Th´evenin model and determine the parameter values along with their uncertainties. The Bayesian posterior distributions are obtained by sampling chains using the Metropolis algorithm. Unlike classical numerical methods, which require repeated parameter inference for changes in battery physical properties, the NNSs offer a fast, differentiable, and scalable alternative and allow efficient updating of parameters as batteries age. This enables fast gradient-based inference and real-time applications in battery management and control. To sum up, NNSs offer a lightweight and fast alternative for modelling and parameterising batteries, providing significant advantages over traditional numerical approaches once trained, despite the initial training time.
Neural Network Surrogates for Ultra-fast Inference in Lithium-ion Battery Models
Ebrahimpour, B. (Author). 4 Jun 2025
Student thesis: Doctoral Thesis